Tanishjain9/floor-segmentation-dataset

Indoor Floor Segmentation Dataset

Custom indoor RGB image dataset for semantic floor segmentation, built for training lightweight models such as Fast-SCNN for real-time, edge/robotics perception applications.

The dataset targets detection of the traversable floor region from indoor RGB images.

---

Dataset Description

This dataset contains RGB indoor images paired with pixel-level semantic segmentation masks. Only one foreground class is manually annotated — all remaining pixels are treated as background.

• Task: Binary semantic segmentation (Floor vs. Background)
• Domain: Indoor scenes
• Intended use: Real-time floor/traversability segmentation for robotics and edge deployment (e.g. ONNX / OpenVINO / TensorRT pipelines)
• License: MIT

Class Labels

Class ID	Class
0	Background
1	Floor

The objective is to segment the visible floor region while treating all non-floor regions — furniture, walls, objects, shadows, clutter, etc. — as background.

---

Dataset Structure

floor/
├── train/
│   ├── images/
│   └── masks/
│
├── val/
│   ├── images/
│   └── masks/
│
└── test/
    ├── images/
    └── masks/

Each image in images/ has a corresponding pixel-aligned mask in masks/ with the same filename. Masks are single-channel, with pixel values corresponding to the class IDs above.

Data Fields

Field	Description
image	RGB indoor photograph
mask	Single-channel segmentation mask (0 = Background, 1 = Floor)

---

Dataset Splits

Split	Images	Purpose
Train	7,128	Model training
Validation	681	Model selection and evaluation
Test	681	Final evaluation

Total images: 8,490

---

Baseline Model Results

A Fast-SCNN model trained on this dataset achieved the following on the held-out validation set:

Metric	Score
Pixel Accuracy	94.9%
mIoU	87.8%

The trained model was subsequently exported to ONNX and OpenVINO IR for edge/CPU deployment. See the accompanying Fast-SCNN Floor Segmentation project for the full training and deployment pipeline.

---

Dataset Limitations

This is a custom indoor floor segmentation dataset and may not represent all possible indoor environments. Model performance trained on this data may vary with:

• Different floor materials
• Unseen environments
• Different lighting conditions
• Shadows and reflections
• Significant camera viewpoint changes
• Cluttered indoor scenes
• Objects with floor-like appearance (e.g. beds, blankets, low furniture)

The dataset contains a single manually annotated foreground class and is therefore focused specifically on floor segmentation, rather than general indoor scene understanding or multi-class obstacle classification.

---

Citation

If you use this dataset, please cite this repository.

@dataset{indoor_floor_segmentation,
  title  = {Indoor Floor Segmentation Dataset},
  author = {Tanish Jain},
  year   = {2026},
  note   = {Custom dataset for real-time indoor floor semantic segmentation}
}

License

This dataset is released under the MIT License.